ACT-R and SOAR

Overview

ACT-R and SOAR are the two most influential cognitive architectures in cognitive science. ACT-R emphasizes memory activation mechanisms and production rule matching, while SOAR emphasizes problem-space search and automatic compilation of experience. Both attempt to build a "unified theory of cognition" but take fundamentally different paths.

---

1. ACT-R (Adaptive Control of Thought -- Rational)

1.1 Basic Architecture

ACT-R has been developed by John Anderson (Carnegie Mellon University) since 1976, with the current version being ACT-R 7.x.

graph TD
    subgraph ACT-R Architecture
        PM[Perceptual Modules<br/>Visual, Aural] --> BUF[Buffers]
        BUF --> PS[Production System<br/>Pattern Matching]
        DM[Declarative Memory] --> BUF
        PS --> DM
        PS --> MM[Motor Modules<br/>Manual, Vocal]
        PS --> GM[Goal Module<br/>Goal Buffer]
        GM --> PS
    end

1.2 Core Components

Declarative Memory

Stores factual knowledge in basic units called chunks:

chunk: addition-fact-3-4
  isa: addition-fact
  addend1: 3
  addend2: 4
  sum: 7

Each chunk has an activation value \(A_i\) that determines retrieval speed and probability:

\[ A_i = B_i + \sum_j W_j \cdot S_{ji} + \epsilon_i \]

where:

• \(B_i\): Base-Level Activation -- reflects usage frequency and recency
• \(W_j \cdot S_{ji}\): Spreading Activation -- associative strength from current context
• \(\epsilon_i\): Noise term (follows a Logistic distribution)

Base-Level Activation Computation:

\[ B_i = \ln \left( \sum_{j=1}^{n} t_j^{-d} \right) \]

where \(t_j\) is the time interval from the \(j\)-th use to the present, and \(d\) is the decay parameter (typically \(d \approx 0.5\)).

Power-Law Forgetting

The base-level activation formula embodies the power-law decay of human memory -- memories that are older and less frequently used are harder to retrieve, which closely matches psychological experimental data.

Retrieval Probability:

\[ P(\text{retrieve}_i) = \frac{1}{1 + e^{-(A_i - \tau)/s}} \]

where \(\tau\) is the retrieval threshold and \(s\) is the noise parameter.

Retrieval Latency:

\[ T_i = F \cdot e^{-f \cdot A_i} \]

The higher the activation value, the faster the retrieval.

Procedural Memory

Stores skill-based knowledge as production rules:

Production: retrieve-addition
  IF  Goal buffer contains (isa: add, num1: =x, num2: =y, result: nil)
  THEN Retrieve from declarative memory (isa: addition-fact, addend1: =x, addend2: =y)
       Place result in retrieval buffer

Each production rule also has a utility value, updated via reinforcement learning:

\[ U_i(n) = U_i(n-1) + \alpha [R_i(n) - U_i(n-1)] \]

where \(R_i\) is the reward received and \(\alpha\) is the learning rate.

Buffers

Buffers serve as the communication interface between modules. Each buffer can hold only one chunk at a time (extremely limited capacity), corresponding to the capacity limitations of human working memory.

Buffer	Corresponding Module	Function
Goal	Goal module	Current goal and task context
Retrieval	Declarative memory	Memory retrieval results
Visual	Visual module	Visual attention focus
Manual	Motor module	Motor control commands
Imaginal	Imaginal module	Temporary manipulation of problem representations

1.3 ACT-R Cognitive Cycle

Each cycle (~50ms):

1. Pattern Matching: Check which production rules' conditions are satisfied by current buffer contents
2. Conflict Resolution: Select the best production based on utility values
3. Execution: Fire the selected production's action part (request retrieval, set buffers, issue motor commands, etc.)

---

2. SOAR (State, Operator And Result)

2.1 Basic Architecture

SOAR has been developed by Allen Newell, John Laird, and Paul Rosenbloom since 1983, with the current version being Soar 9.x.

graph TD
    subgraph SOAR Architecture
        WM[Working Memory] --> DM2[Decision Procedure<br/>Decision Cycle]
        LTM[Long-Term Memory] --> WM
        LTM --> PM2[Procedural<br/>Production Rules]
        LTM --> SM[Semantic]
        LTM --> EM[Episodic]
        DM2 --> OP[Operator Application]
        OP --> WM
        DM2 -->|Impasse| IMP[Subgoal Creation<br/>Impasse & Subgoaling]
        IMP -->|Resolution| CH[Chunking<br/>Experience Compilation]
        CH --> PM2
    end

2.2 Core Concepts

Problem Space

SOAR models all cognitive activity as search in a problem space:

\[ \text{Problem Space} = \langle S, O, I, G \rangle \]

• \(S\): Set of states
• \(O\): Set of operators (state transition functions)
• \(I\): Initial state
• \(G\): Goal state (or goal test function)

Decision Cycle

Each SOAR decision cycle consists of five phases:

1. Input: Acquire new information from perception into working memory
2. Propose: Production rules propose available operators
3. Evaluate: Production rules compare and evaluate candidate operators
4. Decision: Select the best operator
5. Apply: Apply the selected operator, modifying working memory

Impasse and Subgoaling

When the decision process cannot continue (e.g., no available operators, multiple indistinguishable operators), SOAR automatically creates a subgoal to resolve the impasse:

Impasse Type	Description	Example
Tie	Multiple operators cannot be distinguished	Two approaches look equally good
Conflict	Multiple operators contradict each other	One rule suggests forward, another suggests backward
No-change	No available operators	Don't know what to do
Constraint-failure	Selected operator cannot be applied	Preconditions not met

2.3 Chunking: Automatic Compilation of Experience

Chunking is SOAR's most fundamental learning mechanism. When a subgoal is resolved, SOAR automatically compiles the resolution process into a new production rule:

\[ \text{Chunk} = \text{Conditions}(\text{impasse context}) \rightarrow \text{Actions}(\text{impasse resolution result}) \]

Process:

1. Reasoning within the subgoal resolves the impasse
2. SOAR traces back to analyze: which working memory elements led to this result?
3. These elements become conditions, and the result becomes the action, creating a new production rule
4. Next time the same situation is encountered, the new rule matches directly without creating a subgoal

Analogy: Like a human needing to derive a math problem the first time, but after repeated practice being able to directly "see" the answer.

2.4 Long-Term Memory Types (Soar 9+)

Memory Type	Contents	Access Method
Procedural	Production rules	Automatic matching (parallel)
Semantic	Facts and concepts (graph structure)	Query retrieval
Episodic	Snapshots of past experiences	Timeline or content query

Cross-Reference

For a detailed discussion of memory systems, see Episodic and Semantic Memory.

---

3. ACT-R vs. SOAR Comparison

3.1 Core Differences

Dimension	ACT-R	SOAR
Theoretical goal	Cognitive model (simulating humans)	General intelligence (AGI)
Core mechanism	Activation + Utility	Search + Chunking
Memory	Activation-driven retrieval	Production matching + Semantic/Episodic
Learning	Activation adjustment + Utility learning	Chunking + RL
Parallelism	Buffer-level parallelism	Production-level parallelism
Temporal modeling	Precise time predictions (ms-level)	No emphasis on temporal prediction
Application domains	Psychology experiment simulation	Game AI, robotics, military
Programming language	Lisp	C++/Java/Python

3.2 Comparison on the Same Problem

Task: Compute 3 + 4

ACT-R Approach:

1. Goal buffer: (add num1:3 num2:4 result:nil)
2. Production match: retrieve-addition fires
3. Declarative memory retrieval: addition-fact-3-4 (highest activation)
4. Retrieval success: sum = 7
5. Production match: store-result fires
6. Goal buffer: (add num1:3 num2:4 result:7)

SOAR Approach:

1. State: (add ^num1 3 ^num2 4 ^result nil)
2. Propose operators: recall-sum, count-up, count-down
3. Evaluate: recall-sum has highest priority
4. Apply recall-sum: match known fact 3+4=7
5. State update: (add ^num1 3 ^num2 4 ^result 7)

If no direct memory exists (triggering an impasse), SOAR creates a subgoal, solves it via a counting strategy, then learns via Chunking:

Subgoal: Solve via count-up strategy
Chunk: IF (add ^num1 3 ^num2 4) THEN (^result 7)

---

4. Connections to Modern LLM Agents

4.1 Concept Mapping

Classic Concept	LLM Agent Counterpart
ACT-R Declarative Memory	Vector database + RAG
ACT-R Activation Values	Embedding similarity + Temporal decay
ACT-R Production Rules	Tool definitions + System prompts
SOAR Problem-Space Search	Tree of Thoughts
SOAR Impasse → Subgoal	"I'm not sure, let me break down this problem"
SOAR Chunking	Experience summaries written to memory (Reflexion)

4.2 Insights

1. ACT-R's activation mechanism: Can be used to design memory retrieval priorities for LLM agents
2. SOAR's Chunking: Analogous to compiling experience into verbalized rules in Reflexion
3. ACT-R's temporal predictions: Can be used to model response latency of LLM agents
4. SOAR's subgoaling: Analogous to task decomposition and recursive planning in LLM agents

---

5. Other Important Cognitive Architectures

Architecture	Core Feature
CLARION	Dual-process theory (explicit + implicit knowledge)
Icarus	Hierarchical concepts and skills
LIDA	Global Workspace Theory
Sigma	Graphical models unifying cognitive functions
OpenCog	Open-source AGI cognitive architecture

---

Summary

ACT-R and SOAR represent two major traditions in cognitive architecture research: ACT-R focuses more on precise modeling of human cognition, while SOAR focuses more on general problem-solving capability. The core ideas of both -- activation mechanisms, production-based reasoning, problem-space search, and experience compilation -- continue in new forms in the LLM era.

---

References

1. Anderson, J.R. (2007). How Can the Human Mind Occur in the Physical Universe? Oxford University Press.
2. Anderson, J.R. et al. (2004). An Integrated Theory of the Mind. Psychological Review, 111(4), 1036-1060.
3. Laird, J.E. (2012). The Soar Cognitive Architecture. MIT Press.
4. Newell, A. (1990). Unified Theories of Cognition. Harvard University Press.
5. Laird, J.E., Newell, A. & Rosenbloom, P.S. (1987). SOAR: An Architecture for General Intelligence. Artificial Intelligence, 33(1), 1-64.

---

评论